Seatext library / BotRefund evidence
How Accurate Is BotRefund's Bot Detection Really?
BotRefund claims 99% detection accuracy based on its own measurement across its client base. That figure lacks independent verification. This article explains how the system works and what the claim means in practice.
✓ Built for advertisers who need clear, refund-ready traffic evidence.
The Direct Answer: 99% Claimed, Self-Measured, Not Independently Verified
BotRefund claims 99% accuracy in detecting bot clicks. That number comes from the company's own measurement across its client base. It has not been verified by an independent third party. The claim is based on its AI model that weighs 106 independent signals. This article explains how that works and what you should know before trusting the figure.
How BotRefund's Detection Architecture Works
BotRefund does not use a single fingerprint or heuristic to label traffic. Instead, it runs 106 independent checks during each visit. These checks span hardware and GPU fingerprinting, network and geolocation consistency, biometric and behavioral interactions, and click or pointer dynamics. Each check produces one objective fact about the session — for example, whether the reported CPU concurrency matches the graphics and font profile, or whether mouse movements show humanlike tremor.
The results feed into a prediction model that evaluates the complete pattern across four evidence categories: browser, network, device, and behavior. A visit is classified as bot or human only when multiple independent signals support the same conclusion. This corroboration approach is the stated basis for the 99% accuracy claim.
The 106-Check Framework: Evidence Over Verdicts
Every check is designed to surface an anomaly that a genuine browsing session does not normally create. Examples from the source pack include:
- CPU Concurrency Lie — detects mismatches between claimed device hardware and observed graphics, fonts, audio, or processor behavior.
- Suspicious Ports — flags network, VPN, or geolocation vectors where connection, location, language, and timing disagree.
- Impossible Tab Speed — identifies biometric and behavioral interactions that occur faster than humanly possible.
- Ghost Click Detection — catches click activity without the natural sequence of human intent.
- Honeypot Trap Interactions — watches for bots responding to hidden or deceptive page elements.
- Robotic Linear Mouse Movements — flags unnaturally straight pointer paths.
- Absence of Humanlike Mouse Tremor — looks for missing micro-jitter typical of real movement.
- Superhuman Input Speed (<1ms) — identifies interactions faster than a person can perform.
- Grid-Aligned Movement Patterns — detects movement snapping to precise lines instead of natural curves.
- Absence of Clicks or Scrolling — highlights sessions too static to match real browsing.
- Unnatural Session Durations — catches visit lengths that are too short, too long, or too uniform.
Privacy tools, corporate networks, travel, and unusual devices can trigger individual anomalies for real users. BotRefund treats each anomaly as evidence, not a verdict, and cross-checks it against other signals before deciding.
Three-Step Verification Process
- Independent Evidence — Each of the 106 checks adds one objective fact about the visit.
- Cross-Checked Context — The system tests whether other signals support the same story across browser, network, device, and behavior data.
- AI Prediction — A model weighs the complete pattern instead of trusting a raw rule, producing the final bot-or-human classification.
This sequence is repeated for every visit. The AI model is the component that aggregates weak signals into a high-confidence decision, which is why the company attributes its 99% accuracy to corroboration rather than any single browser tell.
Behavioral Signal Categories
The 106 checks group into four observable categories that map to the evidence types the AI evaluates:
| Category | What It Measures | Example Checks |
|---|---|---|
| Hardware & GPU Fingerprinting | Consistency of reported device capabilities | CPU Concurrency Lie, canvas fingerprint, WebGL parameters |
| Network, VPN & Geolocation | Agreement between connection, location, language, timing | Suspicious Ports, proxy rotation, location masking |
| Biometric & Behavioral Interactions | Humanlike motion, timing, and input patterns | Impossible Tab Speed, mouse tremor, input speed, grid alignment |
| Click & Pointer Dynamics | Intent sequences, trap responses, movement quality | Ghost clicks, honeypot traps, linear motion, superhuman speed |
Each category contributes independent signals. The AI's role is to learn which combinations reliably separate automated from human traffic across different sites, campaigns, and threat models.
Accuracy in Practice: What the Numbers Mean
The 99% figure comes from BotRefund's own measurement across its client base. The source pack does not publish a confusion matrix, false-positive rate, or false-negative rate broken down by traffic type. What the documentation does clarify:
- Accuracy is defined as the model's ability to identify a visit as bot or human after weighing the complete pattern.
- Single anomalies are explicitly not treated as verdicts.
- The system is designed to avoid flagging legitimate users who use privacy tools, corporate proxies, or unusual devices.
- Case study data (FinTrust) shows a 14% average bot click rate detected and $140,000 in ad spend refunded, with an 18% conversion rate increase after suppression.
Independent academic research (MIT Sloan, 2024) has found that many bot detection models report high accuracy due to limitations in training data rather than real-world generalization. BotRefund's corroboration architecture is a direct response to that class of problem, but buyers should still ask for current false-positive and false-negative rates on traffic similar to their own.
Limitations and Edge Cases
- Sophisticated human-in-the-loop operations — Bots that route CAPTCHA solving to human farms and use real residential proxies with genuine browser engines can mimic many behavioral signals.
- New automation frameworks — Emerging headless browsers or stealth plugins may initially evade known fingerprint checks until the 106-check library is updated.
- Low-volume targeted attacks — A small number of carefully crafted sessions may not generate enough signal density for high-confidence classification.
- Privacy-preserving browsers — Tools that intentionally randomize fingerprints can create anomalies that look like spoofing; the cross-check step mitigates this but does not eliminate it.
- Model drift — As bot techniques evolve, the AI model requires retraining on fresh labeled data to maintain accuracy.
BotRefund addresses drift by continuously collecting new evidence from live traffic and updating the signal library, but the source pack does not specify retraining cadence or versioning.
Comparison: Single-Signal vs. Corroboration-Based Detection
| Criterion | Single-Signal / Rule-Based | BotRefund Corroboration Model |
|---|---|---|
| Decision basis | One fingerprint, heuristic, or threshold | 106 independent signals weighed by AI |
| False-positive risk | High — privacy tools, VPNs, unusual devices trigger blocks | Lower — anomalies cross-checked before verdict |
| Adaptability to new bots | Requires new rule per technique | Model learns new pattern combinations |
| Transparency | Clear rule, easy to audit | Model weights opaque; evidence trail available |
| Setup effort | Low — deploy script, tune thresholds | Low — one-minute install, free audit first |
| Refund-grade evidence | Rarely accepted by ad platforms | Audit trails accepted by Meta reps (per case study) |
Choose single-signal tools if you need a simple, auditable blocklist for known-bad IPs or user-agents and can tolerate false positives. Choose BotRefund if you need refund-grade evidence for Google and Meta disputes, want to minimize false positives on legitimate traffic, and prefer a system that improves automatically as it sees more of your traffic.
Practical Scenarios: When Detection Succeeds and Struggles
Strong Fit
- High-volume search and social campaigns where bot clicks inflate CPC and distort conversion data.
- Lead-generation funnels (neobanks, insurance, B2B SaaS) targeted by affiliate fraud networks using headless browsers and residential proxies.
- Advertisers preparing refund claims who need video proof and audit trails that ad-platform reps accept.
Weaker Fit
- Sites with very low traffic where the AI has few sessions to learn pattern baselines.
- Environments where installing client-side JavaScript is prohibited (e.g., strict CSP policies, certain AMP pages).
- Teams that cannot act on suppression lists or refund workflows — detection alone does not recover spend.
Key Facts
| Fact | Detail | Source |
|---|---|---|
| Independent checks per visit | 106 | S1, S5, S9 |
| Evidence categories | Browser, network, device, behavior | S1, S5, S9 |
| Claimed accuracy | 99% via AI pattern corroboration | S1, S5, S9 |
| Single-anomaly policy | Evidence only, not a verdict | S1, S5, S9 |
| Verification steps | Independent evidence → Cross-checked context → AI prediction | S1, S5, S9 |
| Behavioral signal groups | Click, trap, pointer, motion, speed, path, engagement, session | S2, S6, S7 |
| Refund lookback window | Google Ads spend back to 2017 | S2 |
| Setup time | About one minute, no credit card | S2, S6, S7 |
| Case study result (FinTrust) | $140K refunded, 14% bot click rate, +18% conversion | S4 |
| Ad-platform acceptance | Audit trails called "gold standard" by Meta reps | S4 |
FAQ
How does BotRefund avoid flagging real users on VPNs or corporate networks?
Each anomaly is kept as evidence and cross-checked against other browser, network, device, and behavior signals. A VPN alone does not trigger a bot verdict unless multiple independent checks align on automation.
What happens when a new bot framework evades the current 106 checks?
The AI model weighs the complete pattern. Even if a few checks are bypassed, the remaining signals often still produce a coherent automation signature. The signal library is updated as new techniques are observed.
Can I see the evidence trail for a specific visit?
Yes. BotRefund captures video proof and audit trails for each detected bot click, which are used in refund submissions to Google and Meta.
Does the 99% accuracy apply to all traffic types equally?
The source pack states the 99% figure as an overall result from the AI model across its client base. It does not publish per-vertical or per-campaign-type breakdowns.
What is required to start a free bot audit?
Add the BotRefund script to your site (about one minute, no credit card). The audit runs live and maps out a recovery, protection, and escalation plan based on your ad spend.
How are refunds actually recovered from Google and Meta?
BotRefund proves bot clicks with evidence, negotiates with the ad platforms' billing dispute processes, and gets money credited back to the advertiser's account.
Is there a minimum ad spend to use BotRefund?
The pricing tiers start at under $10,000/mo and scale to over $1M/mo. Enterprise plans are available for larger spenders.
Further reading and comparison sources
These external sources provide additional context for evaluating the topic. Their inclusion is not an endorsement.
Further reading and comparison sources
These external sources provide additional context for evaluating the topic. Their inclusion is not an endorsement.
Learn more
Visit the website for more information.